Protocol
draft-0
Scoring specified; corpus not qualified
Publication withheld
No ranking without reproducible evidence.
HumanBench has not qualified a comparison of detectors. Earlier unsupported scores are withdrawn. Individual checks performed on another service are not benchmark evidence.
Eligible results
0
No public rank or score
User checks
Excluded
Destination activity is never promoted into ranking data
Detector draft-0 scoring
The scoring core is reproducible, but it cannot qualify a result or emit a rank. Corpus and subject evidence are still missing.
- Decision threshold
- AI probability ≥ 0.50
- Primary metric
- Balanced accuracy
- Uncertainty
- 95% Wilson intervals for sensitivity and specificity; no balanced-accuracy interval claimed
- Missing data
- Explicit, retained in class denominators, and disqualifying
Ranking and tie breaks are disabled in draft-0. A separate versioned amendment and independent review are required before publication.
Publication rule
A comparison remains withheld until every evaluated subject has an exact version and the same protocol can be rerun against a rights-cleared corpus.
- Corpus provenance, rights, exclusions and contamination review
- Exact provider, detector, model or tool version
- Repeatable run and scoring command with retained output digest
- Coverage, uncertainty, tie and missing-data policies